Apple · apple

Apple M2 Ultra

Apple M2 Ultra has 128 GB of unified memory at 819 GB/s — about 89.28 GiB usable after driver and compositor overhead. 2090 of 2118 indexed models fit at 4K context with q8_0 KV. Note only 96 GB of its 128 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
128 GB
LPDDR5-6400
Bandwidth
819 GB/s
1024-bit bus
Tensor FP16
—
dense
TDP
—
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1795vision language 191audio tts 21image 2audio asr 39video 16embedding 26

What fits at 4K context

largest quantization that fits, per model · 2090 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
command-a-plus-05-2026-bf16MoEIQ3_XS219B95.16 GiB0.27 GiB95.98 GiB0.02 GiB29±37%
GLM-4.7MoEIQ2_XS358B94.53 GiB0.76 GiB95.89 GiB0.11 GiB30±37%
GLM-4.7-REAP-218B-A32BMoEIQ3_M218B94.52 GiB0.76 GiB95.87 GiB0.13 GiB25±37%
Laguna-S-2.1MoEQ6_K_L118B94.83 GiB0.17 GiB95.58 GiB0.42 GiB34±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ3_S235B94.50 GiB0.39 GiB95.47 GiB0.53 GiB29±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-TQ1_0363B94.37 GiB0.51 GiB95.47 GiB0.53 GiB28±37%
GLM-4.6-Derestricted-v3MoEIQ2_XS357B94.10 GiB0.76 GiB95.46 GiB0.54 GiB30±37%
GLM-4.6MoEIQ2_XS357B94.10 GiB0.76 GiB95.46 GiB0.54 GiB30±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_S236B94.48 GiB0.39 GiB95.45 GiB0.55 GiB29±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_S236B94.48 GiB0.39 GiB95.45 GiB0.55 GiB29±37%
Qwen3-235B-A22BMoEQ3_K_S235B94.48 GiB0.39 GiB95.45 GiB0.55 GiB29±37%
ERNIE-4.5-300B-A47B-PTUD-IQ2_XXS300B94.32 GiB0.45 GiB95.44 GiB0.56 GiB7±8.3%
MiMo-V2-FlashMoEKV unresolvedUD-IQ2_XXS310B94.58 GiB0.25 GiB95.43 GiB0.57 GiB36±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ3_S236B94.70 GiB0.14 GiB95.43 GiB0.57 GiB35±37%
DeepSeek-V2.5MoEIQ3_S236B94.70 GiB0.14 GiB95.43 GiB0.57 GiB35±37%
DeepSeek-Coder-V2-InstructMoEIQ3_S236B94.70 GiB0.14 GiB95.43 GiB0.57 GiB35±37%
MiniMax-M2.7MoEUD-Q3_K_M229B94.29 GiB0.51 GiB95.34 GiB0.66 GiB35±37%
Behemoth-X-123B-v2Q6_K123B93.68 GiB0.73 GiB95.11 GiB0.89 GiB7±8.3%
Mistral-Large-Instruct-2411Q6_K123B93.68 GiB0.73 GiB95.11 GiB0.89 GiB7±8.3%
gemma-4-26B-A4B-it-Uncensored-MAXMoEF3225.8B94.02 GiB0.24 GiB94.79 GiB1.21 GiB7±8.3%
grok-2MoEQ2_K_L270B93.41 GiB0.53 GiB94.64 GiB1.36 GiB13±37%
MiniMax-M3MoEIQ1_M427B93.82 GiB0.25 GiB94.63 GiB1.37 GiB36±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q6_K121B92.51 GiB1.34 GiB94.43 GiB1.57 GiB7±8.3%
Mixtral-8x22B-Instruct-v0.1MoEQ5_K_M141B93.11 GiB0.46 GiB94.18 GiB1.82 GiB13±37%
Mixtral-8x22B-v0.1MoEQ5_K_M141B93.11 GiB0.46 GiB94.18 GiB1.82 GiB13±37%
Mixtral-8x22B-v0.1MoEQ5_K_M141B93.10 GiB0.46 GiB94.18 GiB1.82 GiB13±37%
MiniMax-M2.1MoEI1-IQ3_M229B93.13 GiB0.51 GiB94.18 GiB1.82 GiB36±37%
MiniMax-M2.5MoEI1-IQ3_M229B93.13 GiB0.51 GiB94.18 GiB1.82 GiB36±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q6_K123B93.42 GiB0.05 GiB94.05 GiB1.95 GiB37±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedTQ1_0402B92.59 GiB0.40 GiB93.57 GiB2.43 GiB44±37%
DeepSeek-V4-FlashMoEQ2_K291B92.86 GiB0.03 GiB93.49 GiB2.51 GiB40±37%
GLM-4.6-REAP-268B-A32BMoEQ2_K_L269B92.11 GiB0.76 GiB93.46 GiB2.54 GiB28±37%
MiniMax-M2MoEQ3_K_S229B92.31 GiB0.51 GiB93.36 GiB2.64 GiB36±37%
GLM-4.5-Air-DerestrictedMoEQ6_K110B92.37 GiB0.38 GiB93.33 GiB2.67 GiB29±37%
GLM-4.5-AirMoEQ6_K110B92.37 GiB0.38 GiB93.33 GiB2.67 GiB29±37%
Mistral-Small-4-119B-2603MoEUD-Q6_K119B92.60 GiB0.05 GiB93.23 GiB2.77 GiB37±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q5_K_M139B91.98 GiB0.51 GiB93.03 GiB2.97 GiB31±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q5_K_M139B91.98 GiB0.51 GiB93.03 GiB2.97 GiB31±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ3_S229B91.94 GiB0.51 GiB92.99 GiB3.01 GiB36±37%
dots.llm1.instMoEQ4_K_L143B89.95 GiB2.06 GiB92.59 GiB3.41 GiB29±37%
Step-3.7-FlashUD-IQ4_NL201B90.63 GiB1.34 GiB92.56 GiB3.44 GiB7±8.3%
Qwen3.5-397B-A17BMoEIQ1_M403B91.53 GiB0.06 GiB92.19 GiB3.81 GiB42±37%
Kimi-Linear-48B-A3B-InstructMoEBF1649.1B91.54 GiB0.06 GiB92.16 GiB3.84 GiB7±8.3%
GLM-4.5MoEUD-IQ1_S358B90.38 GiB0.76 GiB91.74 GiB4.26 GiB31±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ3_XS235B90.25 GiB0.39 GiB91.22 GiB4.78 GiB30±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ3_XS235B90.25 GiB0.39 GiB91.22 GiB4.78 GiB30±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ5_K_L124B90.28 GiB0.18 GiB91.00 GiB5.00 GiB34±37%
Solar-Open2-250BMoEIQ3_XXS250B89.86 GiB0.40 GiB90.83 GiB5.17 GiB39±37%
MiMo-V2.5MoEKV unresolvedUD-IQ2_M311B89.93 GiB0.25 GiB90.77 GiB5.23 GiB37±37%
Nex-N2-ProMoEIQ1_M397B90.02 GiB0.06 GiB90.68 GiB5.32 GiB42±37%
GLM-4.6VMoEQ6_K_L108B89.58 GiB0.38 GiB90.54 GiB5.46 GiB30±37%
GLM-4.5VMoEI1-Q6_K108B89.28 GiB0.38 GiB90.23 GiB5.77 GiB30±37%
Hermes-4-405BUD-IQ1_M406B88.23 GiB1.05 GiB90.11 GiB5.89 GiB8±8.3%
Trinity-Large-PreviewMoEIQ2_XXS399B88.58 GiB0.50 GiB89.65 GiB6.35 GiB42±37%
Trinity-Large-TrueBaseMoEIQ2_XXS399B88.58 GiB0.50 GiB89.65 GiB6.35 GiB42±37%
gpt-oss-20b-hereticMoEQ5_120.9B88.57 GiB0.06 GiB89.16 GiB6.84 GiB22±37%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ5_120.9B88.57 GiB0.06 GiB89.16 GiB6.84 GiB22±37%
Hermes-3-Llama-3.1-405BIQ1_M406B87.08 GiB1.05 GiB88.95 GiB7.05 GiB8±8.3%
Dolphin3.0-R1-Mistral-24BF3223.6B87.82 GiB0.33 GiB88.82 GiB7.18 GiB8±8.3%
Dolphin3.0-Mistral-24BF3223.6B87.82 GiB0.33 GiB88.82 GiB7.18 GiB8±8.3%
●From the file◐Predictedwhat these mean

Speed is modeled, not measured: decode is memory-bandwidth bound, so tokens per second is bytes read per token against achievable bandwidth. Mixture-of-experts models carry a wider band because only the routed experts are read each step, and few have been measured publicly.

Questions people ask

What AI models can a Apple M2 Ultra run?
2090 of 2118 indexed open-weight models fit a Apple M2 Ultra at 4,096 context with q8_0 KV cache, the largest being command-a-plus-05-2026-bf16 at IQ3_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Apple M2 Ultra actually have?
Its nameplate is 128 GB, but about 89.28 GiB is available to a model once driver and compositor overhead is accounted for, and only 96 GB of the pool can be allocated to the GPU at all.
Is a Apple M2 Ultra fast for local AI?
Its memory bandwidth is 819 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.